SAP AI Core is designed to manage the lifecycle of AI assets, including the fine-tuning of foundation models. The most effective approach for a custom business process is to leverage a pre-trained model from the catalog and adapt it to the specific domain (customer service transcripts) via fine-tuning, which is a core capability of the service. This balances efficiency with customization.

SAP AI Core provides automated scaling based on user-defined metrics (like RPS) and offers different resource plans (e.g., for testing vs. production). This combination allows the service to automatically adjust resources to meet demand, optimizing both performance during peaks and cost during low-usage periods, which is a best practice for cloud-native inference services.

In SAP AI Core, an Execution is the running instance of a workflow. Workflow templates (often written in Argo YAML) define the directed acyclic graph (DAG) of steps, which can include different types of actions (like data processing and model training). This is the native mechanism for orchestrating complex, multi-step pipelines that may involve resources both inside and outside the core AI runtime.

To deploy a model for inference, SAP AI Core needs more than just the weights/parameters. It requires: 1) The model binary itself, 2) A scoring script that defines how to load the model and process inference requests (the `predict` interface), and 3) A specification of the environment (dependencies, Python version) to ensure consistent execution. This package of artifacts allows AI Core to containerize and serve the model.

SAP AI Launchpad provides built-in comparison features for executions. After running evaluation workflows for different model versions, you can select them and use tools to compare key metrics (like MAE, RMSE) side-by-side in charts and tables. This provides a standardized, traceable, and efficient way to make data-driven decisions about model promotion.

Proactive model monitoring requires a closed-loop system. SAP AI Core can be integrated with applications to log inputs and predictions. This data can then be analyzed (potentially using SAP’s monitoring features or custom checks) to detect data drift or performance decay. Based on these metrics, automated workflows in AI Core can be triggered to evaluate the model on new data or initiate retraining, which is the industry best practice for maintaining model health.

An AI Scenario is a fundamental organizational concept in SAP AI Core. It provides a logical grouping mechanism for all artifacts (data, code, models, executions, deployments) that belong to a single business application (e.g., “Customer Sentiment Analysis” or “Spare Parts Forecasting”). This simplifies management, access control, and tracking of the AI lifecycle for that particular use case.

For a script to be runnable in the cloud-agnostic, containerized environment of SAP AI Core, it cannot have hardcoded paths or parameters. AI Core injects configuration (like the location of input datasets from Data Connections, hyperparameters, and where to save the output model) via environment variables or command-line arguments. The script must be designed to read these values to be portable and configurable.

The standard, cloud-native integration pattern is to treat the AI model as a microservice. SAP AI Core provides secure, API-managed inference endpoints. A side-by-side extension (like an SAP BTP application) or a cloud application can call this REST API using secure OAuth 2.0 authentication. This keeps the systems decoupled, allows independent scaling, and leverages the managed service features of AI Core.

SAP AI Core is designed to manage the lifecycle of AI assets, including the fine-tuning of foundation models. The most effective approach for a custom business process is to leverage a pre-trained model from the catalog and adapt it to the specific domain (customer service transcripts) via fine-tuning, which is a core capability of the service. This balances efficiency with customization.0